Google Gemini Flaws Allowed AI-Powered Data Theft

Researchers have disclosed three now-patched vulnerabilities in Google's Gemini AI assistant that could have enabled data theft and privacy breaches. Dubbed the "Gemini Trifecta," the flaws allowed attackers to use indirect prompt injection to steal sensitive user information, including saved data and location. The vulnerabilities affected Gemini Cloud Assist, its Search Personalization Model, and its Browsing Tool. Attackers could inject malicious instructions into logs or a user's search history, which Gemini would later execute. These commands could then force the AI's browsing tool to exfiltrate the stolen data to an attacker-controlled server, demonstrating how AI systems themselves can become attack vectors.